API prices =/= actual inference cost for the company
OuterSpaceBootyHole on
Interesting marketing angle. „Give me $200 to spite me“ is just findom in the BDSM world.
invyros on
I love being a burden to corporations, a net negative on their books.
I jump from free trial to free trial, always using the service as much as possible.
I always pay off my credit cards, earning points and cashback while never paying interest.
I use the free tier at cloud providers to their maximum potential for all of my side projects.
If I were inclined to actually pay for a ChatGPT subscription, it would bring me so much joy to know I could hurt them so much as just an individual.
emirsolinno on
I feel like if we burned this much cash on something else, say eliminating fossil energy, we could achieve something great. But instead we have a fun tool that gives cocaine effect to tech bros and CEOs
Qudit314159 on
> SemiAnalysis has calculated how big that gap really is. After testing subscription tiers from both OpenAI and Anthropic – running long-horizon coding and agentic tasks until weekly limits were exhausted – the firm found that the cost of theoretical maximum usage of these plans if priced at standard API rates far exceeds what users actually pay.
No shit. Everyone knows that subscriptions are cheaper with heavy use. Figuring out what the API cost would be doesn’t tell us anything about OpanAI’s actual expenses.
Impressive_Rent_8162 on
Need to get some Chinese bot farms on the case
Spirited-Sir-3034 on
The most interesting part of the article isn’t the $200 subscription potentially costing OpenAI thousands. It’s that agentic workflows can consume up to 1,000x more tokens than a normal prompt. If AI agents become the default way people work, does the subscription model survive, or do we eventually move back to pay-per-use pricing?
FunkyPete on
The secret is to get TWO $200 ChatGPT subscriptions, and then have each one keep the other 100% busy 24/7.
WTFOMGBBQ on
I max out THREE $200 Claud max accounts every month, I wonder if im costing a antropic $42,000
habeaskoopus on
This is bs to drive more subs. They invented every aspect of their pricing. Including token values and blah blah blah.
KlyptoK on
Does the analysis assume the cost of running the service is the token cost and not the reality of it being a fraction of the API token cost?
Do they mean opportunity cost or the loss of offset for development expenses?
glitterandnails on
With these companies losing tons of money even from basic user use, one must ask the question: “What is the endgame here?”
Helgafjell4Me on
I’ll just keep using the free version, not signed in, until they inevitably start requiring paid subscriptions. It’s not like I use it that much anyways, only for helping me work thru ideas on occasion.
I have paid Gemini at work and I’ve used it a bunch… for learning more about fishing my local water holes, lol…
mista_r0boto on
Time to run local llms instead.
savagepanda on
This means the token charges are too expensive. It includes the training costs and capex costs and with a healthy margin added. They are definitely not optimized for efficiency. Meanwhile a deepseek subscription costs 100X less based on token usage. It’s probably based on a longer depreciation curve and cheaper electricity costs. , but the difference is staggering.
BrianD-mage on
The problem is that your usage data can be used for them to grift more investors into giving them way more money than you’re costing them. Best way to make it go away is to not engage with it.
MinuetInUrsaMajor on
Sounds like the average gym membership…
MacPR on
Lots of llms inches behind chatgtp for a fraction of the cost. Openai can’t afford to raise prices.
postitnote on
API cost isn’t the right way to estimate it though. These providers are able to optimize the agentic use case where you have very high prefix cache hit rates. Prefill (encoding the incoming prompt) is much cheaper from cache.
Cached prefill costs 10% of the price vs uncached. And with the bulk of requests coming with the same large system prompt, you can deduplicate parts of it for everyone.
It is still likely they lose money with these plans, but the estimate done by the article is not accurate. You would think that these AI labs would aggressively optimize these agentic usecases.
sprietsma on
Set up multiple accounts to feed each other prompts non-stop. AI ouroboros
nervez on
how much does a free account use cost them? I assume endless queries coming from non-paying customers would cost them similarly.
No-Sympathy-686 on
Malicious compliance is calling….
Strict-Carrot4783 on
This fucking clown world…
thadowski on
itd still cost your ecosystem
FriendlyKillerCroc on
That is true of almost every single subscription you use so this is more stupid news. The price is literally designed around average use.
Doctor_Amazo on
I’m glad that people are finally noticing that the math isn’t mathing for any gen-ai business.
DrowningKrown on
Is this just an elaborate chatGPT ad lol? Their cost isn’t actually $14K.
Necessary_Regret3329 on
M’lord, I have a cunning plan!
pronult3 on
Duhhh, I guess I better subscribe then.
Nice try marketing department.
Im_gumby_damnit on
So wait, people are paying money to post this complete garbage everywhere on the internet? I don’t think people realize how much AI stifles innovation. It just makes people lazy and ultimately unmotivated to innovate.
Necessary_Solid_9462 on
It seems every tech product is „over sold“, like your „unlimited“ data plan. There will be some „up to“ language in the fine print.
cereal7802 on
Thats why high usage accounts get warned, and banned. AI companies are similar to webhosting companies in the 2010s. advertise unlimited everything, then when anyone tries to use it, point them to terms of use and terminate their account.
FlournoyFlennory on
Ollama with Qwen and about 15 lines of code – instant vibe coding chat my little droogies.
otw on
Yeah this is roughly what we saw at a company I worked at. The company started out by giving devs stipends to buy AI plans when it was new cause it was changing so rapidly and everyone had their own preferences.
We started to settle on a few plans and decided to switch to enterprise/team accounts which did not have a flat rate. All our developers who were previously getting around a $200 stipend a month for AI plans were now costing us ~$15k a month.
We went back to the stipend because there’s really nothing stopping you from doing that it’s not even against their ToS or anything. It’s discouraged because you lose out on some enterprise features, but we mostly weren’t using those anyway.
We then got pushed back into the enterprise plan which I believe was done because an investor of ours wanted us to pay full price. I might be cynical but a lot of times I feel like startups are just a way to funnel money to bigger companies and pump usage. We used to get pressured to use cloud services egregiously (the „startup to AWS“ pipeline) and now it feels like we are being pushed to use tokens egregiously.
IAMERROR1234 on
I don’t use them at all but, that sounds like a „them“ problem.
LevelFix83 on
This is the first time I’ve been interested in the $200 subscription. Maybe they should put this in their advertisements.
terrorTrain on
>
the firm found that the cost of theoretical maximum usage of these plans if priced at standard API rates far exceeds what users actually pay.
This is not the same as costing then 14k. Stupid headline
bootycuddles on
This feels like a trap to use it so they can justify more and more data centers in our backyards to steal our electricity and leave us on the hook for paying it. I’m good.
raincoater on
Nice try ChatGPT.
Leave A Reply
Du musst angemeldet sein, um einen Kommentar abzugeben.
42 Kommentare
The crash is coming.
It’s not a bubble because Uber. 🤪
challenge….accepted.
API prices =/= actual inference cost for the company
Interesting marketing angle. „Give me $200 to spite me“ is just findom in the BDSM world.
I love being a burden to corporations, a net negative on their books.
I jump from free trial to free trial, always using the service as much as possible.
I always pay off my credit cards, earning points and cashback while never paying interest.
I use the free tier at cloud providers to their maximum potential for all of my side projects.
If I were inclined to actually pay for a ChatGPT subscription, it would bring me so much joy to know I could hurt them so much as just an individual.
I feel like if we burned this much cash on something else, say eliminating fossil energy, we could achieve something great. But instead we have a fun tool that gives cocaine effect to tech bros and CEOs
> SemiAnalysis has calculated how big that gap really is. After testing subscription tiers from both OpenAI and Anthropic – running long-horizon coding and agentic tasks until weekly limits were exhausted – the firm found that the cost of theoretical maximum usage of these plans if priced at standard API rates far exceeds what users actually pay.
No shit. Everyone knows that subscriptions are cheaper with heavy use. Figuring out what the API cost would be doesn’t tell us anything about OpanAI’s actual expenses.
Need to get some Chinese bot farms on the case
The most interesting part of the article isn’t the $200 subscription potentially costing OpenAI thousands. It’s that agentic workflows can consume up to 1,000x more tokens than a normal prompt. If AI agents become the default way people work, does the subscription model survive, or do we eventually move back to pay-per-use pricing?
The secret is to get TWO $200 ChatGPT subscriptions, and then have each one keep the other 100% busy 24/7.
I max out THREE $200 Claud max accounts every month, I wonder if im costing a antropic $42,000
This is bs to drive more subs. They invented every aspect of their pricing. Including token values and blah blah blah.
Does the analysis assume the cost of running the service is the token cost and not the reality of it being a fraction of the API token cost?
Do they mean opportunity cost or the loss of offset for development expenses?
With these companies losing tons of money even from basic user use, one must ask the question: “What is the endgame here?”
I’ll just keep using the free version, not signed in, until they inevitably start requiring paid subscriptions. It’s not like I use it that much anyways, only for helping me work thru ideas on occasion.
I have paid Gemini at work and I’ve used it a bunch… for learning more about fishing my local water holes, lol…
Time to run local llms instead.
This means the token charges are too expensive. It includes the training costs and capex costs and with a healthy margin added. They are definitely not optimized for efficiency. Meanwhile a deepseek subscription costs 100X less based on token usage. It’s probably based on a longer depreciation curve and cheaper electricity costs. , but the difference is staggering.
The problem is that your usage data can be used for them to grift more investors into giving them way more money than you’re costing them. Best way to make it go away is to not engage with it.
Sounds like the average gym membership…
Lots of llms inches behind chatgtp for a fraction of the cost. Openai can’t afford to raise prices.
API cost isn’t the right way to estimate it though. These providers are able to optimize the agentic use case where you have very high prefix cache hit rates. Prefill (encoding the incoming prompt) is much cheaper from cache.
Cached prefill costs 10% of the price vs uncached. And with the bulk of requests coming with the same large system prompt, you can deduplicate parts of it for everyone.
It is still likely they lose money with these plans, but the estimate done by the article is not accurate. You would think that these AI labs would aggressively optimize these agentic usecases.
Set up multiple accounts to feed each other prompts non-stop. AI ouroboros
how much does a free account use cost them? I assume endless queries coming from non-paying customers would cost them similarly.
Malicious compliance is calling….
This fucking clown world…
itd still cost your ecosystem
That is true of almost every single subscription you use so this is more stupid news. The price is literally designed around average use.
I’m glad that people are finally noticing that the math isn’t mathing for any gen-ai business.
Is this just an elaborate chatGPT ad lol? Their cost isn’t actually $14K.
M’lord, I have a cunning plan!
Duhhh, I guess I better subscribe then.
Nice try marketing department.
So wait, people are paying money to post this complete garbage everywhere on the internet? I don’t think people realize how much AI stifles innovation. It just makes people lazy and ultimately unmotivated to innovate.
It seems every tech product is „over sold“, like your „unlimited“ data plan. There will be some „up to“ language in the fine print.
Thats why high usage accounts get warned, and banned. AI companies are similar to webhosting companies in the 2010s. advertise unlimited everything, then when anyone tries to use it, point them to terms of use and terminate their account.
Ollama with Qwen and about 15 lines of code – instant vibe coding chat my little droogies.
Yeah this is roughly what we saw at a company I worked at. The company started out by giving devs stipends to buy AI plans when it was new cause it was changing so rapidly and everyone had their own preferences.
We started to settle on a few plans and decided to switch to enterprise/team accounts which did not have a flat rate. All our developers who were previously getting around a $200 stipend a month for AI plans were now costing us ~$15k a month.
We went back to the stipend because there’s really nothing stopping you from doing that it’s not even against their ToS or anything. It’s discouraged because you lose out on some enterprise features, but we mostly weren’t using those anyway.
We then got pushed back into the enterprise plan which I believe was done because an investor of ours wanted us to pay full price. I might be cynical but a lot of times I feel like startups are just a way to funnel money to bigger companies and pump usage. We used to get pressured to use cloud services egregiously (the „startup to AWS“ pipeline) and now it feels like we are being pushed to use tokens egregiously.
I don’t use them at all but, that sounds like a „them“ problem.
This is the first time I’ve been interested in the $200 subscription. Maybe they should put this in their advertisements.
>
the firm found that the cost of theoretical maximum usage of these plans if priced at standard API rates far exceeds what users actually pay.
This is not the same as costing then 14k. Stupid headline
This feels like a trap to use it so they can justify more and more data centers in our backyards to steal our electricity and leave us on the hook for paying it. I’m good.
Nice try ChatGPT.